Re: [Sbcl-help] Garbage collection for server with user sessions

[Nikodemus S. <nik...@ra...>]

On 28 February 2010 13:14, Rudi Schlatte <ru...@co...> wrote:

> I'm not sure it would help at all.  The way I understand the problem, we have 1MB per
> request x 300 requests per second = 300MB/s of memory arriving in the oldest generation.
> This seems to be a pessimal memory usage pattern for our generational gc.  I can't see how
> manually breaking some pointers between objects in oldspace would trigger oldspace
> collection earlier than it does now.  (Remember, gc is a heuristic, as is process

It won't trigger an oldspace collection earlier, no. It may, however
do several other beneficial things: allow some things to be collected
before they are promoted so far (via real pointers or accidental
conservativism); make oldspace scavenging faster (this is conditional
on so many things it's not funny -- but given the right set of
circumstances it can do so.)

> I dimly remember a parameter that sets the number of generations, and setting it to some
> small value (1? 2?) essentially reinstating non-generational behavior -- sorry I can't be
> more specific since I don't have the time to go spelunking in sbcl sources right now.  :(
> But it might be interesting to switch off gengc, just for laughs, and see what happens.

Ooh, that's right! I completely forgot that.

+highest-normal-generation+ and +pseudo-static-generation+ in
src/compiler/generic/parms.lisp

Setting the first to 0 and the second to 1 seems to at least build
just fine -- didn't test properly.

Cheers,

 -- Nikodemus